#!/usr/bin/env python3 """Regenerate every blog PNG from its source PDF in the paper's figure/ tree. The mapping below is the single source of truth for which PDF feeds which PNG. Update both this list and the table in `README.md` when figures are added, removed, or renamed. Usage: cd blog_post/figures python3 regenerate.py Requires PyMuPDF (`pip install pymupdf`). """ import os import sys try: import fitz # PyMuPDF except ImportError: sys.stderr.write( "ERROR: PyMuPDF is required.\n" " Install with: pip install pymupdf\n" ) sys.exit(1) # Folders, relative to this script HERE = os.path.abspath(os.path.dirname(__file__)) REPO = os.path.abspath(os.path.join(HERE, "..", "..")) SRC_ROOT = os.path.join(REPO, "figure") DST = HERE # (src_relative_to_figure/, dst_filename, dpi) FIGS = [ # ---------- Small-dense calibration sweeps (LR / WD / init std x LM / DF) ---------- # These are the upstream "tune-dense-once" sweeps that produce the # reference hyperparameters used throughout the rest of the blog. ("blog_post/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_lr_val_loss_step25k_line_only_dense_ffn_1_only.pdf", "dense_sweep_lm_lr.png", 200), ("blog_post/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_w_decay_val_loss_step25k_line_only_dense_ffn_1_only.pdf", "dense_sweep_lm_wd.png", 200), ("blog_post/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_init_std_val_loss_step25k_line_only_dense_ffn_1_only.pdf", "dense_sweep_lm_init.png", 200), ("blog_post/dim_128_bs256_it100k_moe_sigmoid_activated_experts_scaling_final_dense_ffn_1_only.pdf", "dense_sweep_df_lr.png", 200), ("blog_post/dim_128_bs256_25k_dense_moe_weight_decay_scan_final_avg_loss_dense_ffn_1_only.pdf", "dense_sweep_df_wd.png", 200), ("blog_post/dim_128_bs256_25k_dense_moe_init_std_scan_final_avg_loss_until_3e-1_dense_ffn_1_only.pdf", "dense_sweep_df_init.png", 200), # ---------- Activated-experts sweeps (small-scale Bridge II evidence) ---------- # Four panels matching the paper's fig:activated-experts-scaling layout: # (a) LR sweep, (b) WD sweep, (c) init std sweep, (d) initial training loss # each paired LM (left) / DF (right). ("llms/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_lr_val_loss_step25k_line_only.pdf", "activated_experts_lr_lm.png", 200), ("diffusions/activated_shared_grouped_experts/dim_128_bs256_it100k_moe_sigmoid_activated_experts_scaling_final.pdf", "activated_experts_lr_df.png", 200), ("llms/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_w_decay_val_loss_step25k_line_only.pdf", "activated_experts_wd_lm.png", 200), ("diffusions/df_std_wdecay/dim_128_bs256_25k_dense_moe_weight_decay_scan_final_avg_loss.pdf", "activated_experts_wd_df.png", 200), ("llms/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_init_std_val_loss_step25k_line_only.pdf", "activated_experts_init_lm.png", 200), ("diffusions/df_std_wdecay/dim_128_bs256_25k_dense_moe_init_std_scan_final_avg_loss_until_3e-1.pdf", "activated_experts_init_df.png", 200), ("llms/gen7_text_dim_128_layer_32_dense_ffn_1_8_moe_sigmoid_activated_experts_lr_initial_iterations_2_seed_avg_lr_1e-3_max_step_100.pdf", "activated_experts_initial_loss_lm.png", 200), ("diffusions/activated_shared_grouped_experts/dim_128_bs256_it100k_moe_sigmoid_activated_experts_scaling_initial_iterations_4_seed_avg_lr_1p6e-3_max_step_100.pdf", "activated_experts_initial_loss_df.png", 200), # ---------- MoE architectural-axis LR sweeps (capacity / granularity / # shared / grouped / depth / width) — show that LR optima # persist across MoE architectural variants ---------- ("llms/gen7_text_dim_128_layer_32_dense_ffn_1_4_moe_sigmoid_capacity_lr_val_loss_step25k_line_only.pdf", "lr_sweep_capacity_lm.png", 200), ("diffusions/MoE_capacity/dim_128_bs256_it100k_moe_sigmoid_capacity_final.pdf", "lr_sweep_capacity_df.png", 200), ("llms/gen7_text_dim_128_layer_32_dense_ffn_1_4_moe_sigmoid_granularity_lr_val_loss_step25k_line_only.pdf", "lr_sweep_granularity_lm.png", 200), ("diffusions/granularity/dim_128_bs256_it100k_moe_sigmoid_granularity_final_with_ffn4.pdf", "lr_sweep_granularity_df.png", 200), ("llms/gen7_text_dim_128_layer_32_ffn_1_moe_sigmoid_shared_expert_lr_val_loss_step25k_line_only.pdf", "lr_sweep_shared_lm.png", 200), ("diffusions/activated_shared_grouped_experts/dim_128_bs256_it100k_moe_shared_expert_scan_final.pdf", "lr_sweep_shared_df.png", 200), ("llms/gen7_text_dim_128_layer_32_ffn_1_moe_sigmoid_group_global_norm_lr_val_loss_step25k_line_only.pdf", "lr_sweep_grouped_lm.png", 200), ("diffusions/activated_shared_grouped_experts/dim_128_bs256_it100k_moe_gn_grouped_experts_scaling_final.pdf", "lr_sweep_grouped_df.png", 200), ("llms/gen7_text_dim_128_layer_4_8_16_32_ffn_1_moe_16e4a_model_layer_lr_val_loss_step25k_line_only.pdf", "lr_sweep_depth_lm.png", 200), ("diffusions/depth_width/dim_128_bs256_it100k_layer_lr_scaling_final_ft.pdf", "lr_sweep_depth_df.png", 200), ("llms/gen7_text_dim_128_256_512_1024_layer_32_ffn_1_moe_16e4a_model_width_lr_val_loss_step25k_line_only.pdf", "lr_sweep_width_lm.png", 200), ("diffusions/depth_width/dim_128_bs256_it100k_moe_width_scaling_final_ft.pdf", "lr_sweep_width_df.png", 200), # ---------- Fixed-LR scaling grid (4 axes x 2 modalities) ---------- ("paper_loss_scaling/fixed_lr_llm_activated_experts_scaling.pdf", "fixed_lr_llm_activated.png", 200), ("paper_loss_scaling/fixed_lr_diffusion_activated_experts_scaling.pdf", "fixed_lr_df_activated.png", 200), ("paper_loss_scaling/fixed_lr_llm_capacity_scaling.pdf", "fixed_lr_llm_capacity.png", 200), ("paper_loss_scaling/fixed_lr_diffusion_capacity_scaling.pdf", "fixed_lr_df_capacity.png", 200), ("paper_loss_scaling/fixed_lr_llm_granularity_scaling.pdf", "fixed_lr_llm_granularity.png", 200), ("paper_loss_scaling/fixed_lr_diffusion_granularity_scaling.pdf", "fixed_lr_df_granularity.png", 200), ("paper_loss_scaling/fixed_lr_llm_layer_scaling.pdf", "fixed_lr_llm_layers.png", 200), ("paper_loss_scaling/fixed_lr_diffusion_layer_scaling.pdf", "fixed_lr_df_layers.png", 200), # ---------- Large-scale multi-modal diffusion (4 losses + 2 speedups, # matching the paper's fig:large-scale-complete-mue diffusion # rows: 256P / 512P / 240P-keyframe / 240P-5s-video) ---------- ("diffusions/large_run/gr6_conv_mini_512p_dense_vs_moe_multiresolution_1_256_256_train_loss.pdf", "large_loss_256p.png", 200), ("diffusions/large_run/gr6_conv_mini_512p_dense_vs_moe_multiresolution_1_512_512_train_loss.pdf", "large_loss_512p.png", 200), ("diffusions/large_run/gr6_conv_mini_512p_dense_vs_moe_multiresolution_4_240_432_train_loss.pdf", "large_loss_240p_keyframe.png", 200), ("diffusions/large_run/gr6_conv_mini_512p_dense_vs_moe_multiresolution_57_240_432_train_loss.pdf", "large_loss_240p_video.png", 200), ("diffusions/large_run/gr6_conv_mini_512p_dense_vs_moe_multiresolution_1_256_256_convergence_speedup_vs_dense_step.pdf", "large_speedup_256p.png", 200), ("diffusions/large_run/gr6_conv_mini_512p_dense_vs_moe_multiresolution_57_240_432_convergence_speedup_vs_dense_step.pdf", "large_speedup_240p_video.png", 200), # ---------- Large-scale LLM trio ---------- ("llms/large_run/gen7_text_100b_dense_vs_moe_smoothed_train_loss.pdf", "large_llm_train_loss.png", 200), ("llms/large_run/gen7_text_100b_dense_vs_moe_val_loss.pdf", "large_llm_val_loss.png", 200), ("llms/large_run/gen7_text_100b_dense_vs_moe_convergence_speedup_vs_dense_step.pdf", "large_llm_speedup.png", 200), ] def export(src_rel, dst_name, dpi): """Render the first page of `src_rel` to a PNG at the given DPI.""" src = os.path.join(SRC_ROOT, src_rel) if not os.path.exists(src): print(f" MISSING {src_rel}", file=sys.stderr) return False doc = fitz.open(src) try: pix = doc[0].get_pixmap(matrix=fitz.Matrix(dpi / 72, dpi / 72), alpha=False) pix.save(os.path.join(DST, dst_name)) print(f" OK {dst_name:36s} ({pix.width}x{pix.height} @ {dpi} DPI)") return True finally: doc.close() def main(): print(f"Source root: {SRC_ROOT}") print(f"Destination: {DST}") print(f"Exporting {len(FIGS)} figures...\n") ok = sum(1 for src, dst, dpi in FIGS if export(src, dst, dpi)) missing = len(FIGS) - ok print(f"\n{ok}/{len(FIGS)} regenerated, {missing} missing") return 0 if missing == 0 else 1 if __name__ == "__main__": sys.exit(main())